Deep learning for emergency ascites diagnosis using ultrasonography images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Diagnostic Ultrasound

Background:

• Abdominal free fluid detection is crucial for emergency diagnosis and treatment.
• The Focused Assessment with Sonography for Trauma (FAST) technique is vital in emergency settings.
• Deep learning (DL) offers potential for automated image analysis in medical diagnostics.

Purpose of the Study:

• To investigate the efficacy of deep learning (DL) for identifying peritoneal free fluid in ultrasonography (US) images.
• To develop and evaluate a U-net model for automated ascites detection.
• To assess the utility of DL in supporting inexperienced physicians and non-professionals in emergency diagnoses using the FAST technique.

Main Methods:

• Collected 2985 US images from ascites patients.
• Categorized images into Ascites-1, Ascites-2, and Ascites-3 based on anatomy.
• Trained a U-net model (encoder-decoder architecture) on 90% of the data and tested on the remaining 10%.

Main Results:

• The U-net model achieved high sensitivity and specificity for Ascites-1 (94.38%, 68.13%) and Ascites-2 (97.12%, 86.33%).
• Accuracy and AUC for Ascites-2 were 91.73% and 0.91, respectively.
• Average Dice coefficients were 0.65 (SD=0.21) for Ascites-1 and 0.79 (SD=0.14) for Ascites-2.

Conclusions:

• Deep learning, specifically the U-net model, demonstrates viability for automated ascites diagnosis.
• The proposed DL technique can significantly enhance preliminary diagnoses in emergency scenarios.
• This approach holds potential value for improving the FAST technique in clinical practice.